| The extraction and representation of crowd motion features is the key step in the detection of abnormal behavior.Two different algorithms are used to describe the crowd behavior in this article.The first algorithm is to combine the intensity information of the crowd motion with the distribution information of the crowd as the descriptor of the crowd motion.The intensity information of the crowd motion is described by the kinetic energy of the particle in the video,the distribution information of the crowd is described by the particle entropy in the video frame.In order to avoid the interference of illumination changes or other noise factors in the detection stage of crowd behavior,the kinetic energy and particle entropy of the ten successive frames of the video are regarded as the final decision value to compare with a preset threshold.If the kinetic energy and particle entropy of the test video sequences are all greater than the threshold value that the video sequence contains abnormal crowd behavior.In the second algorithm,the streakline flow similarity of the successive video frames is used as the feature descriptor of the video segment.Firstly the video frames are divided into blocks,count the probability distribution of motion similarity in each video block.Secondly the probability distribution of all the video blocks in the video sequence is used as the feature descriptor of the crowd motion.In the training phase,the crowd behavior in the video sequence is labeled and sent to support vector machine.According to the two classification support vector machine,which is obtained in the training phase,the crowd behavior is classified in the detection stage.In order to verify the performance of the algorithms,experiments are carried out to detect abnormal crowd behavior on the public available datasets which are the UMN datasets and PETS2009 datasets.The experimental results show that the algorithms can detect the abnormal crowd behavior effectively and accurately. |
PDF Full Text Request